With the advent of generative AI technologies like ChatGPT, universities are exploring ways to provide secure access to these tools for faculty, staff, and students. The University of California San Diego (UC San Diego) has taken a unique approach by developing TritonGPT, a specialized language model hosted on local infrastructure and optimized for administrative operations. In an interview with key contributors to TritonGPT's development—Brett Pollak, Director of Workplace Technology Services; Adam Tilghman, Analyst/Architect with Academic Technology Services; and Jack Brzezinski, Senior AI Architect—we gain insights into UC San Diego's motivations, challenges, and long-term vision for deploying AI tools within their community.
Why TritonGPT? TritonGPT emerged from UC San Diego's need to streamline administrative tasks currently reliant on commercial generative AI products. The university identified an opportunity to create an alternative that not only serves as a specialized resource for UC San Diego-specific inquiries but also enhances efficiency in addressing a wide array of academic, administrative, financial, and research-related questions. TritonGPT incorporates advanced indexing of key university websites, enabling accurate responses to diverse inquiries.
Innovative Features: TritonGPT is equipped with a Job Description Writer to simplify the creation of job descriptions, reducing time and effort for hiring managers. Future plans include a "Fund Manager Coach" to support fund managers in understanding financial policies and procedures. TritonGPT is currently in its pilot stage and aims to be accessible to all UC San Diego staff and students.
Local Infrastructure Hosting: UC San Diego opted to host TritonGPT on local infrastructure at the San Diego Supercomputer Center (SDSC). This approach offers short-term benefits such as cost control, improved security, and flexibility for fine-tuning. The long-term advantages include enhanced collaboration with SDSC, cost savings compared to cloud services, and customization for complex data interactions.
Challenges and Considerations: Managing community expectations and ensuring accurate responses from the AI model posed challenges during development. Staff training and addressing concerns related to potential errors, misleading information, and bias in large language models were focal points. UC San Diego emphasizes a gradual rollout, gathering feedback, and continuous refinement.
Long-Term Vision and Future AI Services: UC San Diego envisions an evolving portfolio of AI services over the next six months, evaluating and integrating new offerings from enterprise vendors. The university also anticipates supporting generative AI for instructional purposes once clear guidelines are established.
UC San Diego's TritonGPT showcases a university-driven approach to leveraging generative AI, emphasizing customization, security, and a commitment to meeting the specific needs of its community.
More: https://sr.ithaka.org/blog/how-can-universities-create-ai-tools-for-their-communities/
